How can people maintain agency when delegating work to AI?

AI can produce a draft, analysis or recommendation faster than a person could create it from the beginning. That advantage can change the professional's role from doing the work directly to directing and reviewing it.

The risk is subtle. Responsibility may remain with the person even as the tool starts setting the question, selecting the evidence and defining what a good answer looks like. Maintaining agency means keeping meaningful control of those choices.

Delegation changes who performs the work, not who owns the purpose

Human agency is the capacity to form intentions, make choices and act on them. In AI-assisted work, it means the professional remains able to define the purpose, constrain the process, intervene and decide what happens with the result.

Agency is not the same as manually performing every step. People already delegate to colleagues, software and established processes. Good delegation can free attention for higher-value judgement. The difference is that generative AI can return a polished result while leaving gaps in evidence or reasoning difficult to see.

A person can remain formally accountable while losing practical control. If they cannot explain the goal, identify what the AI changed or recover when the output is wrong, “human in the loop” may describe a position rather than meaningful oversight.

Prompting alone leaves important delegation choices implicit

Prompting helps communicate an instruction, but it may begin after crucial decisions have already been left vague. What outcome is needed? Which evidence is authoritative? What must the AI not do? Who owns the final decision? When should the task stop?

If acceptance criteria appear only after the output, fluency can influence them. The professional may decide that a result is “good enough” because it is coherent, rather than because it serves the original purpose.

General instructions to check the output are also limited. Reviewing an entire finished document can be difficult, especially when errors are sparse or compound across several steps. Delegation design should preserve points where correction remains practical.

Practise a human-led delegation contract

Before using AI, write a short delegation contract. Define the task and intended user, the permitted scope, required evidence, prohibited actions, quality criteria, checkpoints and final decision owner. The contract can be brief, but the key choices should be visible.

For consequential work, form an independent problem frame or initial view before seeing the AI response. This reduces the chance that the output becomes the unquestioned starting point. It is unnecessary for every low-risk task, so choose based on consequence and anchoring risk.

Break longer work into reversible stages. Review an outline before a full draft, inspect extracted evidence before synthesis, and pause before any external action. Ask what would trigger revision, rejection or escalation.

After the task, require the learner to explain what the AI contributed, what they changed and why the final result meets the criteria. Include a recovery exercise in which a hidden error is revealed and the learner must trace its effect.

Protect judgement and foundational competence over time

Cognitive offloading is not inherently harmful. Calculators, search tools and templates reduce effort so people can focus elsewhere. The question is whether the offloaded capability will still be needed to frame, check or recover future work.

Early workplace research reports that knowledge workers perceive changes in where they apply critical thinking when using generative AI. Some findings rely on self-report and do not prove that AI causes objective or permanent skill loss. They do justify monitoring what people cease to practise.

Teams can preserve foundational competence through occasional independent tasks, comparison exercises and explanation of AI-assisted work. The right balance depends on the role, consequence and availability of other reviewers.

Organisational design matters too. Clear authority, traceability, override routes and limits on autonomous action support agency. Learning cannot create meaningful control if the workflow gives the person no time, information or permission to intervene.

Agency is maintained when AI reduces production effort while the professional still owns the direction, understands the evidence and can change course.

Example

A product manager uses AI to draft a decision brief. Before delegating, they record the decision question, intended audience, authorised evidence, excluded assumptions and acceptance criteria.

They review the outline before the full draft and compare it with their own problem frame. The AI recommends a feature based on a plausible but unsupported customer need. The product manager rejects the recommendation and keeps the brief focused on verified evidence.

AI accelerates synthesis and drafting, while the professional retains ownership of the question, evidence and decision.

FAQs

• Is all cognitive offloading harmful?

  No. Offloading can reduce routine effort and support better work. Risk increases when people offload the understanding they later need to frame, verify, explain or recover a consequential task.

• Should people always form an independent answer before using AI?

  Not for every low-risk task. An independent frame is more valuable where AI could anchor judgement, evidence is contested or the consequence is material. Use a proportionate rule rather than adding unnecessary duplication.

• How can managers tell whether agency is being maintained?

  Ask who defined the goal and criteria, where the person intervened, what output they rejected and whether they can explain and recover the result. Tool usage or final approval alone provides little evidence.